Spatialworld commited on
Commit
9354f78
·
verified ·
1 Parent(s): 70c04d3

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +246 -250
README.md CHANGED
@@ -1,250 +1,246 @@
1
- ---
2
- language:
3
- - en
4
- license: cc-by-nc-4.0
5
- tags:
6
- - embodied-ai
7
- - vision-language
8
- - spatial-reasoning
9
- - 3d-navigation
10
- - multi-agent
11
- datasets:
12
- - ai2thor
13
- - carla
14
- - procthor
15
- - virtualhome
16
- - embodiedcity
17
- size_categories:
18
- - 1K<n<10K
19
- ---
20
-
21
- # SpatialWorld Benchmark
22
-
23
- **A Multi-Platform Benchmark for Spatial Reasoning and Embodied Task Execution**
24
-
25
- ## 🎯 Overview
26
-
27
- SpatialWorld is a comprehensive benchmark designed to evaluate spatial reasoning and embodied task execution capabilities of Multi-modal Large Language Models (MLLMs) and Vision-Language Models (VLMs). The benchmark spans multiple simulation platforms and diverse task categories.
28
-
29
- ### Key Features
30
-
31
- - **Multi-Platform Coverage**: AI2Thor, CARLA, ProcTHOR, VirtualHome, EmbodiedCity
32
- - **Diverse Tasks**: Navigation, object manipulation, multi-agent coordination, and spatial reasoning
33
- - **Unified Action Space**: Consistent action representation across all platforms
34
- - **Rich Annotations**: Golden actions and success conditions for reproducible evaluation
35
-
36
- ## 📊 Dataset Statistics
37
-
38
- This Hugging Face repository contains the **SpatialWorld Benchmark** with 588 tasks in the unified dataset:
39
-
40
- | Platform | Unified Tasks | Full Dataset (benchmark.zip) | Task Types | Unique Fields |
41
- |----------|---------------|------------------------------|------------|---------------|
42
- | AI2Thor | 343 | 2,500+ | Object manipulation, navigation | Category, Evaluation_Type, Level |
43
- | CARLA | 80 | 80 | Urban navigation, traffic scenarios | executor, image_url, input_modality, origin_location, weather |
44
- | ProcTHOR | 127 | 127 | Indoor navigation, household tasks | scene_index, Category, Evaluation_Type, Level |
45
- | VirtualHome | 38 | 38 | Multi-agent household activities | executor, image_url, input_modality, origin_location, weather |
46
- | **Total** | **588** | **~2,745** | Multi-platform spatial reasoning | **20 fields total** |
47
-
48
- ## 🗂️ Repository Structure
49
-
50
- ```
51
- spatialworld-test.jsonl # Unified dataset (588 tasks, 20 columns)
52
- benchmark.zip # Full original dataset (8695 files, ~1GB)
53
- README.md # This file
54
- ```
55
-
56
- Two versions of the data are provided:
57
-
58
- 1. **`spatialworld-test.jsonl`** (Unified): All 588 tasks in a unified schema with 20 columns. HF Dataset Viewer parses this file.
59
- - Schema differences handled by including all possible fields
60
- - Missing fields use `null`
61
- - Complex nested structures encoded as JSON strings
62
- - Actions converted to Unified Action Space format
63
-
64
- 2. **`benchmark.zip`** (Full Original Dataset): Complete raw task.json files (~1GB, 8695 files).
65
- - Download and unzip to get the original folder structure: `benchmark/ai2thor/`, `benchmark/carla/`, `benchmark/procthor/`, `benchmark/virtualhome/`
66
- - Use this if you need the original JSON format with platform-specific structures
67
-
68
- ## 🎮 Unified Action Space
69
-
70
- All tasks use a standardized action space:
71
-
72
- ### Navigation
73
- - `Move(direction, distance)` - direction: forward/backward/left/right
74
-
75
- ### Viewpoint & Posture
76
- - `Rotate(direction, angle)` - direction: left/right
77
- - `Tilt(direction, angle)` - direction: up/down
78
- - `ChangePosture(pose)` - standing/sitting/lying
79
-
80
- ### Interaction
81
- - `Pick(object)` - Pick up an object
82
- - `Place(target)` - Place held object at target
83
- - `ChangeState(object, state)` - Toggle object state (on/off)
84
- - `Manipulate(object, action)` - Complex manipulation (open/close/clean/slice)
85
-
86
- ### Task Control
87
- - `EndTask(status)` - Terminate task (success/stopped)
88
- - `Communicate(message)` - Agent-to-agent communication
89
-
90
- ## 📁 Task Format
91
-
92
- Each task contains:
93
-
94
- ```json
95
- {
96
- "task_id": "ai2thor00000",
97
- "task_name": "Place object in target",
98
- "instruction": "Natural language task description",
99
- "scene": "FloorPlan17",
100
- "golden_actions": {
101
- "steps": 10,
102
- "actions": [
103
- "Move(forward, 1.0)",
104
- "Rotate(right, 90)",
105
- "Pick(Object)",
106
- "Place(Target)",
107
- "EndTask(success)"
108
- ]
109
- },
110
- "success_conditions": [...],
111
- "max_steps": 50
112
- }
113
- ```
114
-
115
- ## 🔧 Usage
116
-
117
- ### Loading from Hugging Face
118
-
119
- ```python
120
- from datasets import load_dataset
121
- import json
122
-
123
- # Load the full dataset (all 630 tasks, all 20 fields)
124
- dataset = load_dataset("Spatialworld/Spatialworld-bench", split="train")
125
-
126
- # Access a task - all fields are available
127
- task = dataset[0]
128
- print(f"Task: {task['task_id']} ({task['platform']})")
129
- print(f"Instruction: {task['instruction']}")
130
- print(f"Scene: {task['scene']}")
131
-
132
- # Parse JSON-encoded fields
133
- golden_actions = json.loads(task["golden_actions"])
134
- success_conditions = json.loads(task["success_conditions"])
135
- target_objects = json.loads(task["target_object_types"])
136
-
137
- # Platform-specific fields (null if not applicable)
138
- print(f"Category: {task['Category']}") # Only for ai2thor/procthor
139
- print(f"Executor: {task['executor']}") # Only for carla/virtualhome
140
- ```
141
-
142
- ### Dataset Schema
143
-
144
- | Field | Type | Description | Platforms |
145
- |-------|------|-------------|-----------|
146
- | `task_id` | string | Unique identifier | All |
147
- | `task_name` | string | Human-readable name | All |
148
- | `platform` | string | Platform name | All |
149
- | `instruction` | string | Natural language instruction | All |
150
- | `scene` | string | Scene identifier | ai2thor, carla, virtualhome |
151
- | `max_steps` | int | Maximum steps | All |
152
- | `golden_actions` | string (JSON) | Action sequence | All |
153
- | `success_conditions` | string (JSON) | Success criteria | All |
154
- | `target_object_types` | string (JSON) | Target objects | ai2thor, procthor |
155
- | `success_logic` | string | AND/OR logic | ai2thor, procthor |
156
- | `target_description` | string | Detailed description | ai2thor, procthor |
157
- | `Category` | string | Task category | ai2thor, procthor |
158
- | `Evaluation_Type` | string | Evaluation type | ai2thor, procthor |
159
- | `Level` | string | Difficulty level | ai2thor, procthor |
160
- | `executor` | string | Executor type | carla, virtualhome |
161
- | `image_url` | string | Image path | carla, virtualhome |
162
- | `input_modality` | string | Input type | carla, virtualhome |
163
- | `origin_location` | bool | Origin flag | carla, virtualhome |
164
- | `scene_index` | int | Scene number | procthor |
165
- | `weather` | string | Weather condition | carla, virtualhome |
166
-
167
- ### Loading Original Tasks (Local)
168
-
169
- For full benchmark with original JSON structures:
170
-
171
- ```python
172
- import json
173
- from pathlib import Path
174
-
175
- def load_task(platform, task_id):
176
- task_path = Path(f"benchmark/{platform}/tasks/{task_id}/task.json")
177
- with open(task_path, 'r') as f:
178
- return json.load(f)
179
-
180
- # Example
181
- task = load_task("ai2thor", "ai2thor00000")
182
- print(task["instruction"])
183
- print(task["golden_actions"]["actions"])
184
- ```
185
-
186
- ### Data Format
187
-
188
- The Hugging Face dataset provides a unified schema across all platforms:
189
-
190
- | Field | Type | Description |
191
- |-------|------|-------------|
192
- | `task_id` | string | Unique identifier (e.g., "ai2thor00000") |
193
- | `platform` | string | Platform name (ai2thor/carla/procthor/virtualhome) |
194
- | `task_name` | string | Human-readable task name |
195
- | `instruction` | string | Natural language instruction |
196
- | `scene` | string | Scene identifier (FloorPlan, Town, etc.) |
197
- | `max_steps` | int | Maximum allowed steps |
198
- | `golden_actions_json` | string | JSON-encoded golden action sequence |
199
- | `success_conditions_json` | string | JSON-encoded success conditions |
200
- | `target_object_types_json` | string | JSON-encoded target objects |
201
- | `success_logic` | string | Logic for combining success conditions (AND/OR) |
202
- | `target_description` | string | Detailed target description |
203
- | `platform_specific_json` | string | Platform-specific fields (scene_index, executor, etc.) |
204
-
205
- ### Action Parsing
206
-
207
- ```python
208
- import re
209
-
210
- def parse_action(action_str):
211
- """Parse action string to (action_name, args)"""
212
- match = re.match(r'^(\w+)\(([^)]*)\)$', action_str)
213
- if match:
214
- name = match.group(1)
215
- args = [arg.strip() for arg in match.group(2).split(',')]
216
- return name, args
217
- return None, None
218
-
219
- # Example
220
- action = "Move(forward, 1.0)"
221
- name, args = parse_action(action)
222
- # name: "Move", args: ["forward", "1.0"]
223
- ```
224
-
225
- ## 📏 Evaluation
226
-
227
- Tasks are evaluated based on:
228
-
229
- 1. **Success Rate**: Percentage of tasks completed successfully
230
- 2. **Action Efficiency**: Steps used vs. golden actions
231
- 3. **Goal Achievement**: Satisfaction of success conditions
232
-
233
- ### Success Conditions
234
-
235
- - `object_state`: Target object in desired state
236
- - `object_in_receptacle`: Object placed in correct container
237
- - `polygon_area`: Agent reached target location
238
- - `agent_near_object`: Agent within distance of target
239
-
240
- ## 📜 License
241
-
242
- This dataset is released under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International).
243
-
244
- ## 📦 Sample Data
245
-
246
- A small sample Parquet file (`data/spatialworld_samples.parquet`) containing 2 representative tasks from each platform is provided to enable the Hugging Face Dataset Viewer. The full benchmark is available in the `benchmark/` directory.
247
-
248
- ---
249
-
250
- **Note**: This is an anonymized version of the dataset prepared for peer review.
 
1
+ ---
2
+ language:
3
+ - en
4
+ license: cc-by-nc-4.0
5
+ tags:
6
+ - vision-language
7
+ - spatial-reasoning
8
+ - 3d-navigation
9
+ - multi-agent
10
+ datasets:
11
+ - ai2thor
12
+ - carla
13
+ - procthor
14
+ - virtualhome
15
+ - EmbodiedCity
16
+ size_categories:
17
+ - 1K<n<10K
18
+ ---
19
+
20
+ # SpatialWorld Benchmark
21
+
22
+ **A Multi-Platform Benchmark for Spatial Reasoning and Spatial Task Execution**
23
+
24
+ ## 🎯 Overview
25
+
26
+ SpatialWorld is a comprehensive benchmark designed to evaluate spatial reasoning and spatial task execution capabilities of Multi-modal Large Language Models (MLLMs) and Vision-Language Models (VLMs). The benchmark spans multiple simulation platforms and diverse task categories.
27
+
28
+ ### Key Features
29
+
30
+ - **Multi-Platform Coverage**: AI2Thor, CARLA, ProcTHOR, VirtualHome, EmbodiedCity
31
+ - **Diverse Tasks**: Navigation, object manipulation, multi-agent coordination, and spatial reasoning
32
+ - **Unified Action Space**: Consistent action representation across all platforms
33
+ - **Rich Annotations**: Golden actions and success conditions for reproducible evaluation
34
+
35
+ ## 📊 Dataset Statistics
36
+
37
+ This Hugging Face repository contains the **SpatialWorld Benchmark** with 588 tasks in the unified dataset:
38
+
39
+ | Platform | Unified Tasks | Full Dataset (benchmark.zip) | Task Types | Unique Fields |
40
+ |----------|---------------|------------------------------|------------|---------------|
41
+ | AI2Thor | 343 | 2,500+ | Object manipulation, navigation | Category, Evaluation_Type, Level |
42
+ | CARLA | 80 | 80 | Urban navigation, traffic scenarios | executor, image_url, input_modality, origin_location, weather |
43
+ | ProcTHOR | 127 | 127 | Indoor navigation, household tasks | scene_index, Category, Evaluation_Type, Level |
44
+ | VirtualHome | 38 | 38 | Multi-agent household activities | executor, image_url, input_modality, origin_location, weather |
45
+ | **Total** | **588** | **~2,745** | Multi-platform spatial reasoning | **20 fields total** |
46
+
47
+ ## 🗂️ Repository Structure
48
+
49
+ ```
50
+ spatialworld-test.jsonl # Unified dataset (588 tasks, 20 columns)
51
+ benchmark.zip # Full original dataset (8695 files, ~1GB)
52
+ README.md # This file
53
+ ```
54
+
55
+ Two versions of the data are provided:
56
+
57
+ 1. **`spatialworld-test.jsonl`** (Unified): All 588 tasks in a unified schema with 20 columns. HF Dataset Viewer parses this file.
58
+ - Schema differences handled by including all possible fields
59
+ - Missing fields use `null`
60
+ - Complex nested structures encoded as JSON strings
61
+ - Actions converted to Unified Action Space format
62
+
63
+ 2. **`benchmark.zip`** (Full Original Dataset): Complete raw task.json files (~1GB, 8695 files).
64
+ - Download and unzip to get the original folder structure: `benchmark/ai2thor/`, `benchmark/carla/`, `benchmark/procthor/`, `benchmark/virtualhome/`
65
+ - Use this if you need the original JSON format with platform-specific structures
66
+
67
+ ## 🎮 Unified Action Space
68
+
69
+ All tasks use a standardized action space:
70
+
71
+ ### Navigation
72
+ - `Move(direction, distance)` - direction: forward/backward/left/right
73
+
74
+ ### Viewpoint & Posture
75
+ - `Rotate(direction, angle)` - direction: left/right
76
+ - `Tilt(direction, angle)` - direction: up/down
77
+ - `ChangePosture(pose)` - standing/sitting/lying
78
+
79
+ ### Interaction
80
+ - `Pick(object)` - Pick up an object
81
+ - `Place(target)` - Place held object at target
82
+ - `ChangeState(object, state)` - Toggle object state (on/off)
83
+ - `Manipulate(object, action)` - Complex manipulation (open/close/clean/slice)
84
+
85
+ ### Task Control
86
+ - `EndTask(status)` - Terminate task (success/stopped)
87
+ - `Communicate(message)` - Agent-to-agent communication
88
+
89
+ ## 📁 Task Format
90
+
91
+ Each task contains:
92
+
93
+ ```json
94
+ {
95
+ "task_id": "ai2thor00000",
96
+ "task_name": "Place object in target",
97
+ "instruction": "Natural language task description",
98
+ "scene": "FloorPlan17",
99
+ "golden_actions": {
100
+ "steps": 10,
101
+ "actions": [
102
+ "Move(forward, 1.0)",
103
+ "Rotate(right, 90)",
104
+ "Pick(Object)",
105
+ "Place(Target)",
106
+ "EndTask(success)"
107
+ ]
108
+ },
109
+ "success_conditions": [...],
110
+ "max_steps": 50
111
+ }
112
+ ```
113
+
114
+ ## 🔧 Usage
115
+
116
+ ### Loading from Hugging Face
117
+
118
+ ```python
119
+ from datasets import load_dataset
120
+ import json
121
+
122
+ # Load the full dataset (all 630 tasks, all 20 fields)
123
+ dataset = load_dataset("Spatialworld/Spatialworld-bench", split="train")
124
+
125
+ # Access a task - all fields are available
126
+ task = dataset[0]
127
+ print(f"Task: {task['task_id']} ({task['platform']})")
128
+ print(f"Instruction: {task['instruction']}")
129
+ print(f"Scene: {task['scene']}")
130
+
131
+ # Parse JSON-encoded fields
132
+ golden_actions = json.loads(task["golden_actions"])
133
+ success_conditions = json.loads(task["success_conditions"])
134
+ target_objects = json.loads(task["target_object_types"])
135
+
136
+ # Platform-specific fields (null if not applicable)
137
+ print(f"Category: {task['Category']}") # Only for ai2thor/procthor
138
+ print(f"Executor: {task['executor']}") # Only for carla/virtualhome
139
+ ```
140
+
141
+ ### Dataset Schema
142
+
143
+ | Field | Type | Description | Platforms |
144
+ |-------|------|-------------|-----------|
145
+ | `task_id` | string | Unique identifier | All |
146
+ | `task_name` | string | Human-readable name | All |
147
+ | `platform` | string | Platform name | All |
148
+ | `instruction` | string | Natural language instruction | All |
149
+ | `scene` | string | Scene identifier | ai2thor, carla, virtualhome |
150
+ | `max_steps` | int | Maximum steps | All |
151
+ | `golden_actions` | string (JSON) | Action sequence | All |
152
+ | `success_conditions` | string (JSON) | Success criteria | All |
153
+ | `target_object_types` | string (JSON) | Target objects | ai2thor, procthor |
154
+ | `success_logic` | string | AND/OR logic | ai2thor, procthor |
155
+ | `target_description` | string | Detailed description | ai2thor, procthor |
156
+ | `Category` | string | Task category | ai2thor, procthor |
157
+ | `Evaluation_Type` | string | Evaluation type | ai2thor, procthor |
158
+ | `Level` | string | Difficulty level | ai2thor, procthor |
159
+ | `executor` | string | Executor type | carla, virtualhome |
160
+ | `image_url` | string | Image path | carla, virtualhome |
161
+ | `input_modality` | string | Input type | carla, virtualhome |
162
+ | `origin_location` | bool | Origin flag | carla, virtualhome |
163
+ | `scene_index` | int | Scene number | procthor |
164
+ | `weather` | string | Weather condition | carla, virtualhome |
165
+
166
+ ### Loading Original Tasks (Local)
167
+
168
+ For full benchmark with original JSON structures:
169
+
170
+ ```python
171
+ import json
172
+ from pathlib import Path
173
+
174
+ def load_task(platform, task_id):
175
+ task_path = Path(f"benchmark/{platform}/tasks/{task_id}/task.json")
176
+ with open(task_path, 'r') as f:
177
+ return json.load(f)
178
+
179
+ # Example
180
+ task = load_task("ai2thor", "ai2thor00000")
181
+ print(task["instruction"])
182
+ print(task["golden_actions"]["actions"])
183
+ ```
184
+
185
+ ### Data Format
186
+
187
+ The Hugging Face dataset provides a unified schema across all platforms:
188
+
189
+ | Field | Type | Description |
190
+ |-------|------|-------------|
191
+ | `task_id` | string | Unique identifier (e.g., "ai2thor00000") |
192
+ | `platform` | string | Platform name (ai2thor/carla/procthor/virtualhome) |
193
+ | `task_name` | string | Human-readable task name |
194
+ | `instruction` | string | Natural language instruction |
195
+ | `scene` | string | Scene identifier (FloorPlan, Town, etc.) |
196
+ | `max_steps` | int | Maximum allowed steps |
197
+ | `golden_actions_json` | string | JSON-encoded golden action sequence |
198
+ | `success_conditions_json` | string | JSON-encoded success conditions |
199
+ | `target_object_types_json` | string | JSON-encoded target objects |
200
+ | `success_logic` | string | Logic for combining success conditions (AND/OR) |
201
+ | `target_description` | string | Detailed target description |
202
+ | `platform_specific_json` | string | Platform-specific fields (scene_index, executor, etc.) |
203
+
204
+ ### Action Parsing
205
+
206
+ ```python
207
+ import re
208
+
209
+ def parse_action(action_str):
210
+ """Parse action string to (action_name, args)"""
211
+ match = re.match(r'^(\w+)\(([^)]*)\)$', action_str)
212
+ if match:
213
+ name = match.group(1)
214
+ args = [arg.strip() for arg in match.group(2).split(',')]
215
+ return name, args
216
+ return None, None
217
+
218
+ # Example
219
+ action = "Move(forward, 1.0)"
220
+ name, args = parse_action(action)
221
+ # name: "Move", args: ["forward", "1.0"]
222
+ ```
223
+
224
+ ## 📏 Evaluation
225
+
226
+ Tasks are evaluated based on:
227
+
228
+ 1. **Success Rate**: Percentage of tasks completed successfully
229
+ 2. **Action Efficiency**: Steps used vs. golden actions
230
+ 3. **Goal Achievement**: Satisfaction of success conditions
231
+
232
+ ### Success Conditions
233
+
234
+ - `object_state`: Target object in desired state
235
+ - `object_in_receptacle`: Object placed in correct container
236
+ - `polygon_area`: Agent reached target location
237
+ - `agent_near_object`: Agent within distance of target
238
+
239
+ ## 📜 License
240
+
241
+ This dataset is released under CC BY-NC 4.0 (Creative Commons Attribution-NonCommercial 4.0 International).
242
+
243
+
244
+ ---
245
+
246
+ **Note**: This is an anonymized version of the dataset prepared for peer review.